Res-MoCoDiff: residual-guided diffusion models for motion artifact correction in brain MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Motion artifacts (ARTs) in brain MRI degrade image quality.
• Conventional ART correction methods are time-consuming and burdensome.
• There is a need for efficient and robust ART correction techniques.

Purpose of the Study:

• Introduce Res-MoCoDiff, an efficient denoising diffusion probabilistic model for MRI motion ART correction.
• Evaluate the performance of Res-MoCoDiff against existing methods.
• Demonstrate the clinical potential of Res-MoCoDiff.

Main Methods:

• Developed Res-MoCoDiff using a novel residual error shifting mechanism and a U-net backbone with Swin Transformer blocks.
• Integrated a combined ℓ1+ℓ2 loss function for image sharpness and reduced pixel errors.
• Evaluated on in-silico and in-vivo datasets, comparing with CycleGAN, Pix2pix, and ViT-based diffusion models.

Main Results:

• Res-MoCoDiff demonstrated superior performance in removing motion ARTs across all distortion levels.
• Achieved highest SSIM and lowest NMSE, with PSNR up to 41.91±2.94 dB.
• Reduced average sampling time to 0.37s per batch, a significant improvement over conventional methods.

Conclusions:

• Res-MoCoDiff provides a robust and efficient solution for MRI motion ART correction.
• The method preserves fine structural details while reducing computational overhead.
• Res-MoCoDiff shows potential for integration into clinical workflows to enhance diagnostic accuracy.